search

OPTML-Group / Unlearn-Saliency

[ICLR24 (Spotlight)] "SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation" by Chongyu Fan*, Jiancheng Liu*, Yihua Zhang, Eric Wong, Dennis Wei, Sijia Liu
https://www.optml-group.com/posts/salun_iclr24
MIT License
97 stars 13 forks source link

Question Regarding DDPM Class-Wise Unlearning Iterations #27

Open soyoung2022 opened 5 hours ago

soyoung2022 commented 5 hours ago

I read your paper and am currently trying to reproduce the experimental results. However, I have some questions, so I’m posting this issue.

In the paper regarding the DDPM class-wise unlearning experiments, the appendix mentions that 80,000 iterations are used during retraining. When following the sequence of model training → mask generation → unlearning, how many iterations are used for the initial model training? Additionally, in the configs/cifar10_train.yml file, the n_iters in the training section is set to 800,000. Is this correct?

Thank you for your assistance!

soyoung2022 commented 5 hours ago

So, I understand that the retrain iterations are 80,000, and for the initial model training before unlearning, it is 800,000, is this correct?